We show that the right reasoning recipe can substantially improve the zero-shot performance of existing robot foundation models without new robot data or foundation-scale training. ARC combines action-grounded reasoning traces that explain why an action is taken and what it should cause, a pipeline that generates such traces automatically from existing demonstrations (ARC-Trace-DROID), and model-specific adaptation for both VLAs such as π_0.5 and WAMs such as Cosmos3-Nano-Policy. The adapted models set a new state of the art on RoboLab-120 and MolmoSpaces, and ARC raises π_0.5's real-robot success by over 80 percentage points.

GeomVLA is a Vision-Language-Action model that keeps perception, scene-motion prediction, and action generation in one shared robot-centric 3D frame. It lifts pretrained VLM features into 3D scene tokens and uses a 3D Scene Trajectory Denoiser whose intermediate motion tokens condition a geometry-aware, flow-based action denoiser. GeomVLA achieves state-of-the-art results on CALVIN, is competitive on LIBERO and RoboTwin2.0, and outperforms strong baselines on real robots without robot-action pretraining.

Standard VLAs process every input on a single synchronous clock, which oversamples slow modalities and caps control at the slowest effective rate. DAM-VLA decouples temporal processing per modality with latent buffers refreshed at each sensor's rate, and adds high-frequency modalities through gated cross-attention that leaves the pretrained backbone intact. On seven contact-rich real-world tasks, it more than doubles the success rate of the strongest synchronous baseline while sustaining smooth 100 Hz control.

We systematically analyze VLA design decisions for small and efficient VLAs. Our findings lead us to introduce FLOWER, a 950M-parameter Vision-Language-Action (VLA) policy that achieves state-of-the-art performance across 190 tasks in 10 benchmarks while requiring only ~1% of the pretraining compute of models like OpenVLA. FLOWER introduces intermediate-modality fusion and action-specific Global-AdaLN conditioning to achieve strong performance with improved efficiency. Our approach democratizes VLA development by making high-performance robotic foundation models accessible with commodity hardware, requiring significantly less GPU memory to run.

We introduce BEAST, a B-spline–based action tokenizer that efficiently represents continuous robot actions for generalist policies while maintaining smooth trajectories essential for robot control. BEAST enables efficient action representation and improved performance in VLA models by leveraging the mathematical properties of B-splines for smooth, continuous control. It flexibly supports continuous tokens and discrete tokenization. Experiments across various benchmarks verify good compression with strong performance and smooth behavior without additional temporal aggregation.

We present ReCAT, a language-conditioned policy with structured recurrent memory for tasks that require recalling earlier cues, counting repeated events, or tracking elapsed time. A Mamba-2–based memory with a causal attention layer summarizes the observation stream, and a flow-matching Transformer decoder attends to current and historical features in every block. ReCAT reaches 95.3% on LIBERO, leads on most RMBench tasks, and clearly outperforms short-history baselines on real-robot recall, counting, and timing tasks.

We present PointMapPolicy, a multi-modal imitation learning method that conditions diffusion policies on structured grids of points without downsampling. Our approach fuses point maps with RGB using an xLSTM backbone, enabling direct application of computer vision techniques to 3D data while preserving fine-grained geometric details. On RoboCasa and CALVIN, plus real-robot evaluations, we achieve state-of-the-art performance across diverse manipulation tasks.

We propose Mixture-of-Denoising Experts (MoDE), a generalist policy for guided behavior generation that outperforms dense transformer-based diffusion policies in accuracy, parameter count, and efficiency. Our routing strategy conditions expert selection on the current noise level of the diffusion process. On four imitation learning benchmarks (including CALVIN and LIBERO), MoDE consistently exceeds dense transformer baselines. We pretrain MoDE on a subset of OXE for just 3 days on 6 GPUs and surpass OpenVLA and Octo on SIMPLER. MoDE achieves higher average performance with ~90% fewer FLOPs, ~20% faster inference, and ~40% fewer parameters compared to dense transformer diffusion policies.

We present a diffusion policy for learning from uncurated, reward-free offline data with sparse language labels. Multimodal Diffusion Transformer (MDT) learns complex, long-horizon behaviors and sets a new state of the art on CALVIN. MDT leverages pre-trained vision and language foundation models and aligns multimodal goal specifications in the transformer encoder’s latent space, using two self-supervised auxiliary objectives to better follow goals specified in language and images.

We present BESO, a policy representation for goal-conditioned imitation learning using score-based diffusion models. BESO effectively learns goal-directed, multimodal behavior from uncurated, reward-free offline data and achieves state of the art with as few as three denoising steps.

We introduce the Information Maximizing Curriculum to address mode-averaging in imitation learning by enabling specialization on representable data. A mixture-of-experts policy focuses on different data subsets, with a maximum-entropy objective for full dataset coverage.

DUET-DINO is a latent world model that jointly predicts future side- and wrist-camera representations through cross-view conditioning, enabling latent planning over the full 7-DoF end-effector action space. Trained from scratch on DROID and RoboArena, it outperforms single-view and independent dual-view baselines on reach, angled-reach, and grasp-and-lift tasks. We also find that DINOv3 features capture fine-grained action effects in the wrist view better than V-JEPA 2.

SPARC automatically labels robot demonstrations with structured spatial annotations (bounding boxes, object trajectories, and manipulation phases) and attaches a calibrated reliability score to each label by exploiting the spatio-temporal structure of robot tasks. On our new IA-Bench, SPARC outperforms detection-only baselines while retaining three times more samples at high precision. Models finetuned on SPARC annotations reach state-of-the-art grounding and pointing results among similarly sized models, and policies trained on them perform better in cluttered real-world scenes.

We introduce NaviTrace, a comprehensive benchmark for evaluating embodied navigation capabilities of Vision-Language Models. NaviTrace provides a standardized evaluation framework across diverse navigation tasks, enabling systematic assessment of VLMs' spatial reasoning and instruction-following abilities in diverse real-world robotic navigation scenarios.

We introduce a method to automatically label uncurated, long-horizon robot teleoperation data at scale in a zero-shot manner without human intervention. We combine pre-trained vision-language foundation models to detect objects, propose tasks, segment tasks in large unlabeled interaction datasets, and train language-conditioned policies on the relabeled data. Initial experiments show that our method enables policies that match those trained with oracle human annotations.

D3IL introduces simulation benchmark environments and datasets tailored for imitation learning, designed to evaluate a model’s ability to learn and replicate diverse, multimodal human behaviors. Environments encompass multiple sub-tasks and object manipulations, providing rich diversity often lacking in other datasets. We also propose practical metrics to quantify behavioral diversity and benchmark state-of-the-art methods.

We present a hybrid model combining a differentiable rigid-body model with a recurrent LSTM to accurately model the inverse dynamics of a robot manipulator. A differentiable formulation of barycentric parameters enables end-to-end training jointly with the residual network while preserving physical consistency. On a Franka Emika Panda, the model enables precise and compliant motion tracking.